福州大学学报（自然科学版）2025，Vol.53Issue(2)：168-176,9.DOI:10.7631/issn.1000-2243.24086
自注意力机制下融合图像和点云的三维目标检测
3D object detection based on self-attention mechanism for image fusion and point clouds
摘要
Abstract
To address the issue of low detection accuracy for smaller targets such as pedestrians and occluded objects in autonomous driving,proposes a 3D object detection algorithm that integrates images and point clouds based on a self-attention mechanism.This algorithm improves upon the F-PointNet network,which is based on raw point cloud data processing.By introducing two layers of Transformer-based self-attention mechanism modules into the point cloud feature extraction network,it manages to balance both global and local features of the point cloud,thereby enhancing the detection accuracy of 3D objects.Additionally,elastic net regularization weight decay terms are introduced into the loss function to enhance the generalization capability of the model and achieve higher-precision convergence.A comparative experiment based on the KITTI dataset demonstrates that after introducing the self-attention mechanism and elastic net regularization,the detection accuracy of pedestrians in simple,moderate,and difficult scenarios increased by 6.47%,6.31%,and 5.61%,respectively,compared to the initial model.Similarly,the detection accuracy of cyclists improved by 15.34%,12.88%,and 11.79%in the respective scenarios.关键词
图像/点云/三维目标检测/自注意力机制Key words
image/point cloud/3D object detection/self-attention mechanism分类
信息技术与安全科学引用本文复制引用
林晓鹏,彭育辉,黄炜,陈文强..自注意力机制下融合图像和点云的三维目标检测[J].福州大学学报（自然科学版）,2025,53(2):168-176,9.基金项目
福建省科技厅引导性基金资助项目(2022H0007) （2022H0007）
福建省自然科学基金资助项目(2021J01559) （2021J01559）
